Rasterizer

Splatwizard integrates multiple rasterizers from prior work, which can be directly imported and used. Each rasterizer closely follows the design in the original codebase to minimize tedious migration work.

Standard Gaussian rasterizer

class splatwizard.rasterizer.gaussian.GaussianRasterizer(raster_settings: GaussianRasterizationSettings)

Standard Gaussian Splatting rasterizer from 3D Gaussian Splatting for Real-Time Radiance Field Rendering.

The raster_setting class is defined as

class splatwizard.rasterizer.gaussian.GaussianRasterizationSettings(
    image_height: int,
    image_width: int,
    tanfovx: float,
    tanfovy: float,
    bg: torch.Tensor,
    scale_modifier: float,
    viewmatrix: torch.Tensor,
    projmatrix: torch.Tensor,
    sh_degree: int,
    campos: torch.Tensor,
    prefiltered: bool,
    debug: bool
)

Rasterizer function is defined as

class GaussianRasterizer(nn.Module):
    ...
    def rasterizer(self,
        means3D: Tensor, 
        means2D: Tensor, 
        opacities: Tensor, 
        shs: Tensor | None, 
        colors_precomp: Tensor | None, 
        scales: Tensor | None, 
        rotations: Tensor| None, 
        cov3D_precomp: Tensor | None
    ):
        ...
        return rendered_image, radii

Standard Gaussian rasterizer with depth

class splatwizard.rasterizer.gs_dr_aa.GSDRAAGaussianRasterizer(raster_settings: GSDRAARasterizationSettings)

Standard Gaussian Splatting rasterizer with inv_depth and anti-aliasing.

The raster_setting class is defined as

class splatwizard.rasterizer.gs_dr_aa.GSDRAARasterizationSettings(
    image_height: int
    image_width: int 
    tanfovx: float
    tanfovy: float
    bg: Tensor
    scale_modifier: float
    viewmatrix: Tensor
    projmatrix: Tensor
    sh_degree: int
    campos: Tensor
    prefiltered: bool
    debug: bool
    antialiasing: bool
)

Rasterizer function is defined as

class GSDRAAGaussianRasterizer(nn.Module):
    ...
    def rasterizer(self,
        means3D: Tensor, 
        means2D: Tensor, 
        opacities: Tensor, 
        shs: Tensor | None, 
        colors_precomp: Tensor | None, 
        scales: Tensor | None, 
        rotations: Tensor | None, 
        cov3D_precomp: Tensor | None
    ):
        ...
        return rendered_image, radii, invdepths

Accelerated Gaussian rasterizer

class splatwizard.rasterizer.accel_gs.AccelGaussianRasterizer(raster_settings: AccelGaussianRasterizationSettings)

Gaussian Splatting rasterizer from Taming 3DGS: High-Quality Radiance Fields with Limited Resources

The raster_setting class is defined as

class splatwizard.rasterizer.accel_gs.AccelGaussianRasterizationSettings(
    image_height: int,
    image_width: int,
    tanfovx: float,
    tanfovy: float,
    bg: torch.Tensor,
    scale_modifier: float,
    viewmatrix: torch.Tensor,
    projmatrix: torch.Tensor,
    sh_degree: int,
    campos: torch.Tensor,
    prefiltered: bool,
    debug: bool
)

Rasterizer function is defined as

class AccelGaussianRasterizer(nn.Module):
    ...
    def rasterizer(self,
        means3D: Tensor, 
        means2D: Tensor, 
        opacities: Tensor, 
        shs: Tensor | None, 
        colors_precomp: Tensor | None, 
        scales: Tensor | None, 
        rotations: Tensor| None, 
        cov3D_precomp: Tensor | None
    ):
        ...
        return rendered_image, radii

Speedy-Splat rasterizer

class splatwizard.rasterizer.speedy.SpeedyGaussianRasterizer(raster_settings: SpeedyGaussianRasterizationSettings)

Gaussian Splatting rasterizer from Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives

The raster_setting class is defined as

class splatwizard.rasterizer.speedy.SpeedyGaussianRasterizationSettings(
    image_height: int,
    image_width: int,
    tanfovx: float,
    tanfovy: float,
    bg: torch.Tensor,
    scale_modifier: float,
    viewmatrix: torch.Tensor,
    projmatrix: torch.Tensor,
    sh_degree: int,
    campos: torch.Tensor,
    prefiltered: bool,
    debug: bool
)

Rasterizer function is defined as

class SpeedyGaussianRasterizer(nn.Module):
    ...
    def rasterizer(self,
        means3D: Tensor, 
        means2D: Tensor, 
        opacities: Tensor,
        scores: Tensor,
        shs: Tensor | None, 
        colors_precomp: Tensor | None, 
        scales: Tensor | None, 
        rotations: Tensor| None, 
        cov3D_precomp: Tensor | None
    ):
        ...
        return rendered_image, radii

Speedy-Splat Tensor Core rasterizer

class splatwizard.rasterizer.speedy_tcgs.SpeedyTCGaussianRasterizer(raster_settings: SpeedyTCGaussianRasterizationSettings)

Gaussian Splatting rasterizer from TC-GS: A Faster Gaussian Splatting Module Utilizing Tensor Cores

The raster_setting class is defined as

class splatwizard.rasterizer.speedy_tcgs.SpeedyTCGaussianRasterizationSettings(
    image_height: int,
    image_width: int,
    tanfovx: float,
    tanfovy: float,
    bg: torch.Tensor,
    scale_modifier: float,
    viewmatrix: torch.Tensor,
    projmatrix: torch.Tensor,
    sh_degree: int,
    campos: torch.Tensor,
    prefiltered: bool,
    debug: bool
)

Rasterizer function is defined as

class SpeedyTCGaussianRasterizer(nn.Module):
    ...
    def rasterizer(self,
        means3D: Tensor, 
        means2D: Tensor, 
        opacities: Tensor,
        scores: Tensor,
        shs: Tensor | None, 
        colors_precomp: Tensor | None, 
        scales: Tensor | None, 
        rotations: Tensor| None, 
        cov3D_precomp: Tensor | None
    ):
        ...
        return rendered_image, radii

FlashGS

class splatwizard.rasterizer.flashgs.FlashGSRasterizer(num_vertex, device, MAX_NUM_RENDERED, MAX_NUM_TILES)

Gaussian Splatting rasterizer from FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering

Note: For FlashGSRasterizer, we recommand to directly use splatwizard.modules.render_mixin.FlashGSRenderMixin, which includes both forward and backward method (using standard GS rasterizer in stages where FlashGS is not applicable).

Rasterizer function is defined as

class FlashGSRasterizer(nn.Module):
    ...
    def forward(self, 
        position: Tensor, 
        shs: Tensor, 
        opacity: Tensor, 
        cov3d: Tensor, 
        camera, 
        bg_color: Tensor
    ):
        ...
        return out_color

2D Gaussian Splatting rasterizer

class splatwizard.rasterizer.surfel_gs.SurfelGaussianRasterizer(raster_settings: SurfelGaussianRasterizationSettings)

Gaussian Splatting rasterizer from 2D Gaussian Splatting for Geometrically Accurate Radiance Fields

The raster_setting class is defined as

class splatwizard.rasterizer.surfel_gs.SurfelGaussianRasterizer(
    image_height: int,
    image_width: int,
    tanfovx: float,
    tanfovy: float,
    bg: torch.Tensor,
    scale_modifier: float,
    viewmatrix: torch.Tensor,
    projmatrix: torch.Tensor,
    sh_degree: int,
    campos: torch.Tensor,
    prefiltered: bool,
    debug: bool
)

Rasterizer function is defined as

class SurfelGaussianRasterizer(nn.Module):
    ...
    def rasterizer(self,
        means3D: Tensor, 
        means2D: Tensor, 
        opacities: Tensor,
        scores: Tensor,
        shs: Tensor | None, 
        colors_precomp: Tensor | None, 
        scales: Tensor | None, 
        rotations: Tensor| None, 
        cov3D_precomp: Tensor | None
    ):
        ...
        return rendered_image, radii, allmap

For detailed usage of allmap, please refer to the implemetation of 2DGS in splatwizard/model_zoo/surfel_gs/model.py